From: Michael Roy Ames (michaelroyames@hotmail.com)
Date: Sat Jul 27 2002 - 13:26:57 MDT
Cliff Stabbert <cps46@earthlink.net> wrote:
>
> Specifically, in arguing that some analogies "are better" than others,
> he [Douglas Hofstadter] argues for a "survivability value"
> -- i.e., some analogies give you
> a better chance of survival, or some ways of analogizing give you a
> better chance of survival than others.
>
> With that in mind, I'm curious [...] whether any work
> has been done to combine Copycat-like projects with neural networks,
> genetic algorithms or something along those lines, i.e., evolving
> Copycat's underlying rules and structures based on some type of
> virtualized environment.
>
Cliff,
Work on the Copycat program/architecture has continued. Google for
'Metacat' and 'James B. Marshall'.
James' dissertation <http://www.cs.pomona.edu/marshall/metacat.pdf> gives a
detailed account of how Copycat has been expanded into Metacat, which now
includes many new abilities outlined by Douglas Hofstadter in "Fluid
Concepts and Creative Analogies", and discussed at FARG. This work has not
implemented the kind of 'rule/structure evolution' you ask about, but it
makes a stab at implementing some necessary functionality needed to produce
such a system. Mechanisms for Self-watching, Episodic-memory, Comparing and
contrasting answers, and handling of arbitrary strings have been added.
What has not yet been programmed is the ability to make-up new, high quality
analogy problems - this was deemed to large a step to be taken this time
around, and 'order of magnitude' more difficult. I suspect that the ability
to expand Metacat's Slipnet rules and structures (the thing you are asking
about) would be another 'order of magnitude' step beyond that.
Michael Roy Ames
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:40 MDT